针对非负矩阵盲信号分离(NMF)用于混合像元分解易陷入局部极小值的不足,将非监督端元提取与盲分解方法相结合,构建了一种基于目标端元修正的混合像元盲分解模型(ATGP-NMF)。ATGP-NMF模型利用非监督正交子空间投影算法(ATGP)和非负最小二乘法(NNLS)获取NMF盲分离的初始值,然后将获得初始目标端元光谱与丰度输入NMF模型,通过迭代运算不断逼近优化目标而得到最终的端元光谱和端元丰度。为了检验模型对于各类数据的有效性和适用性,将ATGP-NMF与传统NMF分别应用于模拟仿真数据、室内控制数据和真实遥感影像3类实验数据进行分析验证。结果表明,ATGP-NMF模型具有较好的适用性,在没有先验信息、先验信息很少,以及纯像元假设不存在情况下都能较好地分解混合像元,且能够更好克服局部极小问题,提高混合像元分解的精度。

Spectral unmixing is an important and challenging task in the field of hyperspectral data analysis. The existing methods of blind unmixing have certain limitations. In this paper, we present a new blind unmixing method, namely the ATGP-NMF algorithm, for hyperspectral imagery. The method is based on the improved target endmember acquired through integration of the Automatic Target Generation Process (ATGP) algorithm and the Non-negative Matrix Factorization (NMF). The Harsanyi-Farrand-Chang (HFC) algorithm was introduced firstly to determine the number of target endmembers. Then the ATGP algorithm and Non-Negative Least Squares (NNLS) were used to obtain the spectra and abundances of the target endmembers, which were then used as initial values for the NMF algorithm to obtain the refined endmembers. Finally, an improved cross correlogram spectral matching method was introduced to match the corresponding land cover type of each endmember. Three different sets of data, namely simulated data, laboratory-controlled spectral experimental data and remote sensing imagery, were used in this study to test the effectiveness and robustness of the proposed method, in comparison with the original NMF algorithm. Results from these experiments show that the ATGP-NMF algorithm can obtain endmembers with high accuracy and it is more robust and efficient than the original NMF algorithm in different situations, regardless of the existence of pure pixels, inter-class diversity, or correlation among the endmembers' spectra. The ATGP-NMF algorithm thus has great potential of application in blind unmixing for hyperspectral imagery.